Scholarlink

<p>Myocardial infarction (MI) localization using 12-lead electrocardiograms (ECGs) presents challenges due to the fixed weighting of leads in traditional multi-feature branch (MFB) networks, which may not align with clinical requirements. To overcome this limitation, we propose a Transformer-based MFB network that solely relies on Transformers. The ECG signals are first denoised using a discrete wavelet transform, followed by R-peak detection via the Pan–Tompkins algorithm. Each heartbeat is segmented by extracting a 651-sample window centered around the R-peak and dividing it into 31 fragments of 21 samples. Within the feature branch, we utilize transformers to capture the dependencies between feature segments, thereby extracting the essential features of MI. Finally, we employ transformers to summarize the lead features across the twelve leads, facilitating the extraction of the correlation between the leads and MI subclasses, which include anterior myocardial infarction, anterior septal myocardial infarction, anterior lateral myocardial infarction, inferior myocardial infarction, and inferior lateral myocardial infarction. Results from the PTB-XL diagnostic electrocardiogram database demonstrate the competitive overall performance of the proposed method, achieving an overall accuracy of 88.19% and an F1 score of 0.8566, showing competitive results with, and in some cases matching, existing state-of-the-art methods. These results highlight the promising applications of this research in the field of intelligent health monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A transformer encoder-based multi-feature branch network for myocardial infarction localization

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Peng Xiong,
  • Shiqi Ma,
  • Wenjun Xie,
  • Jieshuo Zhang,
  • Haiman Du,
  • Jing Zhang,
  • Xiaohu Zhou,
  • Xiuling Liu

摘要

Myocardial infarction (MI) localization using 12-lead electrocardiograms (ECGs) presents challenges due to the fixed weighting of leads in traditional multi-feature branch (MFB) networks, which may not align with clinical requirements. To overcome this limitation, we propose a Transformer-based MFB network that solely relies on Transformers. The ECG signals are first denoised using a discrete wavelet transform, followed by R-peak detection via the Pan–Tompkins algorithm. Each heartbeat is segmented by extracting a 651-sample window centered around the R-peak and dividing it into 31 fragments of 21 samples. Within the feature branch, we utilize transformers to capture the dependencies between feature segments, thereby extracting the essential features of MI. Finally, we employ transformers to summarize the lead features across the twelve leads, facilitating the extraction of the correlation between the leads and MI subclasses, which include anterior myocardial infarction, anterior septal myocardial infarction, anterior lateral myocardial infarction, inferior myocardial infarction, and inferior lateral myocardial infarction. Results from the PTB-XL diagnostic electrocardiogram database demonstrate the competitive overall performance of the proposed method, achieving an overall accuracy of 88.19% and an F1 score of 0.8566, showing competitive results with, and in some cases matching, existing state-of-the-art methods. These results highlight the promising applications of this research in the field of intelligent health monitoring.